CSKV: Training-Efficient Channel Shrinking for KV Cache in Long-Context Scenarios

Luning Wang, Shiyao Li, Xuefei Ning, Zhihang Yuan, Shengen Yan, Guohao Dai, Yu Wang
Proceedings of The 4th NeurIPS Efficient Natural Language and Speech Processing Workshop, PMLR 262:468-484, 2024.

Abstract

Large Language Models (LLMs) have been widely adopted to process long-context tasks. However, the large memory overhead of the key-value (KV) cache poses significant challenges in long-context scenarios. Existing training-free KV cache compression methods typically focus on quantization and token pruning, which have compression limits, and excessive sparsity can lead to severe performance degradation. Other methods design new architectures with less KV overhead but require significant training overhead. To address the above two drawbacks, we further explore the redundancy in the channel dimension and apply an architecture-level design with minor training costs. Therefore, we introduce CSKV, a training-efficient Channel Shrinking technique for KV cache compression: (1) We first analyze the singular value distribution of the KV cache, revealing significant redundancy and compression potential along the channel dimension. Based on this observation, we propose using low-rank decomposition for key and value layers and storing the low-dimension features. (2) To preserve model performance, we introduce a bi-branch KV cache, including a window-based full-precision KV cache and a low-precision compressed KV cache. (3) To reduce the training costs, we minimize the layer-wise reconstruction loss for the compressed KV cache instead of retraining the entire LLMs. Extensive experiments show that CSKV can reduce the memory overhead of the KV cache by 80% while maintaining the model’s long-context capability. Moreover, we show that our method can be seamlessly combined with quantization to further reduce the memory overhead, achieving a compression ratio of up to 95%. Code is available at https://github.com/wln20/CSKV.

Cite this Paper

BibTeX
@InProceedings{pmlr-v262-wang24a, title = {{CSKV}: Training-Efficient Channel Shrinking for {KV} Cache in Long-Context Scenarios}, author = {Wang, Luning and Li, Shiyao and Ning, Xuefei and Yuan, Zhihang and Yan, Shengen and Dai, Guohao and Wang, Yu}, booktitle = {Proceedings of The 4th NeurIPS Efficient Natural Language and Speech Processing Workshop}, pages = {468--484}, year = {2024}, editor = {Rezagholizadeh, Mehdi and Passban, Peyman and Samiee, Soheila and Partovi Nia, Vahid and Cheng, Yu and Deng, Yue and Liu, Qun and Chen, Boxing}, volume = {262}, series = {Proceedings of Machine Learning Research}, month = {14 Dec}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v262/main/assets/wang24a/wang24a.pdf}, url = {https://proceedings.mlr.press/v262/wang24a.html}, abstract = {Large Language Models (LLMs) have been widely adopted to process long-context tasks. However, the large memory overhead of the key-value (KV) cache poses significant challenges in long-context scenarios. Existing training-free KV cache compression methods typically focus on quantization and token pruning, which have compression limits, and excessive sparsity can lead to severe performance degradation. Other methods design new architectures with less KV overhead but require significant training overhead. To address the above two drawbacks, we further explore the redundancy in the channel dimension and apply an architecture-level design with minor training costs. Therefore, we introduce CSKV, a training-efficient Channel Shrinking technique for KV cache compression: (1) We first analyze the singular value distribution of the KV cache, revealing significant redundancy and compression potential along the channel dimension. Based on this observation, we propose using low-rank decomposition for key and value layers and storing the low-dimension features. (2) To preserve model performance, we introduce a bi-branch KV cache, including a window-based full-precision KV cache and a low-precision compressed KV cache. (3) To reduce the training costs, we minimize the layer-wise reconstruction loss for the compressed KV cache instead of retraining the entire LLMs. Extensive experiments show that CSKV can reduce the memory overhead of the KV cache by 80% while maintaining the model’s long-context capability. Moreover, we show that our method can be seamlessly combined with quantization to further reduce the memory overhead, achieving a compression ratio of up to 95%. Code is available at https://github.com/wln20/CSKV.} }
Endnote
%0 Conference Paper %T CSKV: Training-Efficient Channel Shrinking for KV Cache in Long-Context Scenarios %A Luning Wang %A Shiyao Li %A Xuefei Ning %A Zhihang Yuan %A Shengen Yan %A Guohao Dai %A Yu Wang %B Proceedings of The 4th NeurIPS Efficient Natural Language and Speech Processing Workshop %C Proceedings of Machine Learning Research %D 2024 %E Mehdi Rezagholizadeh %E Peyman Passban %E Soheila Samiee %E Vahid Partovi Nia %E Yu Cheng %E Yue Deng %E Qun Liu %E Boxing Chen %F pmlr-v262-wang24a %I PMLR %P 468--484 %U https://proceedings.mlr.press/v262/wang24a.html %V 262 %X Large Language Models (LLMs) have been widely adopted to process long-context tasks. However, the large memory overhead of the key-value (KV) cache poses significant challenges in long-context scenarios. Existing training-free KV cache compression methods typically focus on quantization and token pruning, which have compression limits, and excessive sparsity can lead to severe performance degradation. Other methods design new architectures with less KV overhead but require significant training overhead. To address the above two drawbacks, we further explore the redundancy in the channel dimension and apply an architecture-level design with minor training costs. Therefore, we introduce CSKV, a training-efficient Channel Shrinking technique for KV cache compression: (1) We first analyze the singular value distribution of the KV cache, revealing significant redundancy and compression potential along the channel dimension. Based on this observation, we propose using low-rank decomposition for key and value layers and storing the low-dimension features. (2) To preserve model performance, we introduce a bi-branch KV cache, including a window-based full-precision KV cache and a low-precision compressed KV cache. (3) To reduce the training costs, we minimize the layer-wise reconstruction loss for the compressed KV cache instead of retraining the entire LLMs. Extensive experiments show that CSKV can reduce the memory overhead of the KV cache by 80% while maintaining the model’s long-context capability. Moreover, we show that our method can be seamlessly combined with quantization to further reduce the memory overhead, achieving a compression ratio of up to 95%. Code is available at https://github.com/wln20/CSKV.
APA
Wang, L., Li, S., Ning, X., Yuan, Z., Yan, S., Dai, G. & Wang, Y.. (2024). CSKV: Training-Efficient Channel Shrinking for KV Cache in Long-Context Scenarios. Proceedings of The 4th NeurIPS Efficient Natural Language and Speech Processing Workshop, in Proceedings of Machine Learning Research 262:468-484 Available from https://proceedings.mlr.press/v262/wang24a.html.

Related Material